Mistral’s Large Model 2: Competing with Meta and OpenAI’s Newest Innovations

In the realm of advanced AI technology, developments are happening swiftly. Mistral has just unveiled its latest creation, Large 2, on Wednesday, positioning it as a formidable rival to the most recent offerings from OpenAI and Meta, especially in areas like coding, math, and logical reasoning capabilities.

The unveiling of Mistral Large 2 came hot on the heels of Meta introducing its most advanced open source model to date, Llama 3.1 405b. Mistral touts that Large 2 sets new standards in both performance and affordability for open models, supporting these claims with several benchmarks.

In terms of code creation and mathematics, Large 2 seems to leave Llama 3.1 405B in the dust, achieving this feat with just one-third of the parameters – exactly 123 billion.

Mistral emphasized in a press statement that a significant focus during the training of Large 2 was on reducing the model’s tendency to fabricate plausible but incorrect information, known as hallucinations. The firm stated that Large 2 has been trained to exercise greater discernment in its responses, even to the extent of admitting ignorance where applicable.

Despite being a relatively new player in the AI market, the Paris-based startup Mistral has recently garnered $640 million in its Series B funding round, led by General Catalyst, bringing its valuation to $6 billion. Mistral is quickly establishing itself as a notable competitor with its rapid deployment of cutting-edge AI models.

It’s worth mentioning, though, that Mistral’s innovations, like many in this space, are not fully open source — commercial use requires a license. While it offers more accessibility than some, such as GPT-4o, deploying such advanced models is beyond the capacity of many due to the required expertise and infrastructure. This is even more the case with Llama’s 405 billion parameters.

An area where Mistral Large 2, as well as Meta’s latest Llama 3.1, fall short is in multimodal capabilities — a domain where OpenAI currently leads. Multimodal AI systems, which can process both images and text, are a feature increasingly sought after by emerging startups.

Large 2 boasts a 128,000 token capacity, allowing it to process large amounts of data in a single query (equivalent to about a 300-page book). The model’s new multilingual capabilities are significant, understanding a dozen languages ranging from English to Korean, including 80 programming languages. Mistral asserts that Large 2 also generates more succinct responses compared to existing leading models, known for their verbose outputs.

Mistral Large 2 can be accessed on various platforms like Google Vertex AI, Amazon Bedrock, Azure AI Studio, and IBM watsonx.ai. Furthermore, it’s available on Mistral’s own platform under the moniker “mistral-large-2407” and can be trialed on the company’s ChatGPT alternative, le Chat, for free.